| Oceanic Internal waves are common in marginal seas and continental shelf areas,and their formation mechanism is more complex and random in geographical and temporal distribution,with amplitudes up to 240 meters and peaks up to several hundred kilometers.Internal waves are formed beneath the ocean’s surface,and their propagation process can generate strong currents,endangering marine engineering safety.Internal waves with large amplitudes can cause vertical transport motion of water masses,endangering oil rigs and submarines while also influencing coral reefs,fog layers,marine ecology,and other environments.The accurate detection and identification of oceanic internal waves can aid in the utilization of marine resources and the safety of maritime navigation.Remote sensing images with high spatial resolution and extensive coverage have been widely employed in oceanic internal waves analysis and research.Traditional physics-based algorithms are time and energy intensive and have a limited capacity for generalization.The advancement of deep learning opens up new opportunities for making the most of marine remote sensing data and enhancing internal waves extraction.As a result,this research explores a deep learning-based technique for identifying and detecting oceanic internal waves.This paper gathers 1316 scenes of remote sensing images from the Moderate Resolution Imaging Spectroradiometer(MODIS)in the northern South China Sea during the summer of2019 to 2021 based on the characteristics of internal waves in remote sensing image reconstruction.To offer a clear morphology of internal waves for better visibility,various corrections are made,including geometric correction,Bow-tie correction,and image enhancement.Internal waves are small and medium-scale ocean phenomena that account for a minor percentage of ocean remote sensing photos,therefore they are chopped and filtered before being processed.With the help of the Label Img program,remote sensing photos are manually tagged with the internal waves type and location coordinates,and a dataset of 2076 images from the MODIS ocean is created.The algorithm based on candidate region is explored,and the original approach is enhanced based on the features of oceanic internal waves on remote sensing photos.Res Net50 is used as the network’s feature extraction layer rather than VGG16 in order to deepen the network hierarchy and avoid gradient disappearance and information loss throughout the training phase.Based on this,a pyramid structure of FPN features is introduced to combine multiple scales to improve the accuracy of detection and recognition of internal waves with small horizontal scales.The improved method is validated against the original algorithm,and the findings show that the enhanced algorithm can successfully improve detection accuracy while remaining slow in detection speed.The regression-based technique is investigated and improved.To speed up the algorithm’s convergence,the K-means++ algorithm is chosen in order to obtain a more suitable anchoring framework for oceanic internal waves and to speed up the convergence of the algorithm;the CBAM attention module is added to improve the algorithm’s ability to extract oceanic internal waves;and DIOU is used as the loss function for optimization instead of IOU to improve target localization accuracy.The effectiveness of each point of improvement was tested by ablative studies,and compare the final upgraded method’s performance to that of the original algorithm.Ablation tests are carried out in order to validate the effect of each improvement point and to compare the performance of the final enhanced algorithm to that of the original algorithm.According to the results,the revised algorithm enhances the accuracy and speed of identifying oceanic internal waves.Comparing the detection impact of the advanced algorithm based on candidate region and the advanced algorithm based on regression,it is clear that the detection accuracy and detection rate of the advanced algorithm based on regression are higher.An oceanic internal waves detection system is created based on the enhanced regression algorithm,and users can easily and rapidly recognize oceanic internal waves in remote sensing photos.It has been tested to successfully accomplish the identification task and has some application value. |
